Multi-stage classi cation of images from features and related text

The synergy of textual and visual information in Web documents provides great opportunity for improving the image indexing and searching capabilities of Web image search engines. We explore a new approach for automatically classifying images using image features and related text. In particular, we de ne a multi-stage classi cation system which progressively restricts the perceived class of each image through applications of increasingly specialized classi ers. Furthermore, we exploit the related textual information in a novel process that automatically constructs the training data for the image classi ers. We demonstrate initial results on classifying photographs and graphics from the Web.